Skip to content

Level 2 · Model Registry & Packaging

Level 2 moves from "a tracked experiment" to "a deployable artifact": a real model registry that versions models the way git versions code, a real promotion workflow that gates what actually reaches production, and a real, locally-served inference endpoint that proves a registered model can actually answer a request.

What was run, and what wasn't

This level uses a real local MLflow Model Registry, registering real models from Level 1's real tracked runs, promoting them through real stage transitions, and serving one with a real local inference server. A full managed model-serving platform at production scale was not run as part of this course; the underlying serving mechanism is shown directly and honestly instead.

Prerequisites: Level 1.

Modules

  1. The Model Registry: Versioning Models Like Code: a real model registered, with a real, incrementing version number.
  2. Model Stages: Staging, Production, Archived: a real model moved through real stage transitions, confirmed at each step.
  3. Packaging a Model: Signatures & Environments: a real model signature that catches a real malformed request before it reaches the model.
  4. Model Lineage: Tracing a Model Back to Its Run: a real registered model traced back to the exact real run, data, and code that produced it.
  5. Comparing Model Versions: Is the New One Better?: a real side-by-side comparison deciding whether a new version actually deserves to replace the old one.
  6. Model Approval Workflows: Gating Production: a real gate that blocks an unapproved model from reaching production.
  7. Serving a Model Locally: a Real Inference Endpoint: a real HTTP endpoint, backed by a real registered model, answering a real request.
  8. Input/Output Schemas: Catching Bad Requests Early: a real malformed request rejected before it ever reaches the model.
  9. Rolling Back a Bad Model Version: a real rollback, executed and verified.
  10. Project: Register, Promote & Serve a Real Model: the full real lifecycle, end to end.

After Level 2

Level 3 automates everything built by hand in Levels 1–2 into a real pipeline: training, testing, and deploying a model through a real CI/CD process, not a sequence of manual steps.